src/pcache1.c

   1 /*
   2 ** 2008 November 05
   3 **
   4 ** The author disclaims copyright to this source code.  In place of
   5 ** a legal notice, here is a blessing:
   6 **
   7 **    May you do good and not evil.
   8 **    May you find forgiveness for yourself and forgive others.
   9 **    May you share freely, never taking more than you give.
  10 **
  11 *************************************************************************
  12 **
  13 ** This file implements the default page cache implementation (the
  14 ** sqlite3_pcache interface). It also contains part of the implementation
  15 ** of the SQLITE_CONFIG_PAGECACHE and sqlite3_release_memory() features.
  16 ** If the default page cache implementation is overridden, then neither of
  17 ** these two features are available.
  18 **
  19 ** A Page cache line looks like this:
  20 **
  21 **  -------------------------------------------------------------
  22 **  |  database page content   |  PgHdr1  |  MemPage  |  PgHdr  |
  23 **  -------------------------------------------------------------
  24 **
  25 ** The database page content is up front (so that buffer overreads tend to
  26 ** flow harmlessly into the PgHdr1, MemPage, and PgHdr extensions).   MemPage
  27 ** is the extension added by the btree.c module containing information such
  28 ** as the database page number and how that database page is used.  PgHdr
  29 ** is added by the pcache.c layer and contains information used to keep track
  30 ** of which pages are "dirty".  PgHdr1 is an extension added by this
  31 ** module (pcache1.c).  The PgHdr1 header is a subclass of sqlite3_pcache_page.
  32 ** PgHdr1 contains information needed to look up a page by its page number.
  33 ** The superclass sqlite3_pcache_page.pBuf points to the start of the
  34 ** database page content and sqlite3_pcache_page.pExtra points to PgHdr.
  35 **
  36 ** The size of the extension (MemPage+PgHdr+PgHdr1) can be determined at
  37 ** runtime using sqlite3_config(SQLITE_CONFIG_PCACHE_HDRSZ, &size).  The
  38 ** sizes of the extensions sum to 272 bytes on x64 for 3.8.10, but this
  39 ** size can vary according to architecture, compile-time options, and
  40 ** SQLite library version number.
  41 **
  42 ** Historical note:  It used to be that if the SQLITE_PCACHE_SEPARATE_HEADER
  43 ** was defined, then the page content would be held in a separate memory
  44 ** allocation from the PgHdr1.  This was intended to avoid clownshoe memory
  45 ** allocations.  However, the btree layer needs a small (16-byte) overrun
  46 ** area after the page content buffer.  The header serves as that overrun
  47 ** area.  Therefore SQLITE_PCACHE_SEPARATE_HEADER was discontinued to avoid
  48 ** any possibility of a memory error.
  49 **
  50 ** This module tracks pointers to PgHdr1 objects.  Only pcache.c communicates
  51 ** with this module.  Information is passed back and forth as PgHdr1 pointers.
  52 **
  53 ** The pcache.c and pager.c modules deal pointers to PgHdr objects.
  54 ** The btree.c module deals with pointers to MemPage objects.
  55 **
  56 ** SOURCE OF PAGE CACHE MEMORY:
  57 **
  58 ** Memory for a page might come from any of three sources:
  59 **
  60 **    (1)  The general-purpose memory allocator - sqlite3Malloc()
  61 **    (2)  Global page-cache memory provided using sqlite3_config() with
  62 **         SQLITE_CONFIG_PAGECACHE.
  63 **    (3)  PCache-local bulk allocation.
  64 **
  65 ** The third case is a chunk of heap memory (defaulting to 100 pages worth)
  66 ** that is allocated when the page cache is created.  The size of the local
  67 ** bulk allocation can be adjusted using
  68 **
  69 **     sqlite3_config(SQLITE_CONFIG_PAGECACHE, (void*)0, 0, N).
  70 **
  71 ** If N is positive, then N pages worth of memory are allocated using a single
  72 ** sqlite3Malloc() call and that memory is used for the first N pages allocated.
  73 ** Or if N is negative, then -1024*N bytes of memory are allocated and used
  74 ** for as many pages as can be accommodated.
  75 **
  76 ** Only one of (2) or (3) can be used.  Once the memory available to (2) or
  77 ** (3) is exhausted, subsequent allocations fail over to the general-purpose
  78 ** memory allocator (1).
  79 **
  80 ** Earlier versions of SQLite used only methods (1) and (2).  But experiments
  81 ** show that method (3) with N==100 provides about a 5% performance boost for
  82 ** common workloads.
  83 */
  84 #include "sqliteInt.h"
  85
  86 typedef struct PCache1 PCache1;
  87 typedef struct PgHdr1 PgHdr1;
  88 typedef struct PgFreeslot PgFreeslot;
  89 typedef struct PGroup PGroup;
  90
  91 /*
  92 ** Each cache entry is represented by an instance of the following
  93 ** structure. A buffer of PgHdr1.pCache->szPage bytes is allocated
  94 ** directly before this structure and is used to cache the page content.
  95 **
  96 ** When reading a corrupt database file, it is possible that SQLite might
  97 ** read a few bytes (no more than 16 bytes) past the end of the page buffer.
  98 ** It will only read past the end of the page buffer, never write.  This
  99 ** object is positioned immediately after the page buffer to serve as an
 100 ** overrun area, so that overreads are harmless.
 101 **
 102 ** Variables isBulkLocal and isAnchor were once type "u8". That works,
 103 ** but causes a 2-byte gap in the structure for most architectures (since
 104 ** pointers must be either 4 or 8-byte aligned). As this structure is located
 105 ** in memory directly after the associated page data, if the database is
 106 ** corrupt, code at the b-tree layer may overread the page buffer and
 107 ** read part of this structure before the corruption is detected. This
 108 ** can cause a valgrind error if the uninitialized gap is accessed. Using u16
 109 ** ensures there is no such gap, and therefore no bytes of uninitialized
 110 ** memory in the structure.
 111 **
 112 ** The pLruNext and pLruPrev pointers form a double-linked circular list
 113 ** of all pages that are unpinned.  The PGroup.lru element (which should be
 114 ** the only element on the list with PgHdr1.isAnchor set to 1) forms the
 115 ** beginning and the end of the list.
 116 */
 117 struct PgHdr1 {
 118   sqlite3_pcache_page page; /* Base class. Must be first. pBuf & pExtra */
 119   unsigned int iKey;        /* Key value (page number) */
 120   u16 isBulkLocal;          /* This page from bulk local storage */
 121   u16 isAnchor;             /* This is the PGroup.lru element */
 122   PgHdr1 *pNext;            /* Next in hash table chain */
 123   PCache1 *pCache;          /* Cache that currently owns this page */
 124   PgHdr1 *pLruNext;         /* Next in circular LRU list of unpinned pages */
 125   PgHdr1 *pLruPrev;         /* Previous in LRU list of unpinned pages */
 126                             /* NB: pLruPrev is only valid if pLruNext!=0 */
 127 };
 128
 129 /*
 130 ** A page is pinned if it is not on the LRU list.  To be "pinned" means
 131 ** that the page is in active use and must not be deallocated.
 132 */
 133 #define PAGE_IS_PINNED(p)    ((p)->pLruNext==0)
 134 #define PAGE_IS_UNPINNED(p)  ((p)->pLruNext!=0)
 135
 136 /* Each page cache (or PCache) belongs to a PGroup.  A PGroup is a set
 137 ** of one or more PCaches that are able to recycle each other's unpinned
 138 ** pages when they are under memory pressure.  A PGroup is an instance of
 139 ** the following object.
 140 **
 141 ** This page cache implementation works in one of two modes:
 142 **
 143 **   (1)  Every PCache is the sole member of its own PGroup.  There is
 144 **        one PGroup per PCache.
 145 **
 146 **   (2)  There is a single global PGroup that all PCaches are a member
 147 **        of.
 148 **
 149 ** Mode 1 uses more memory (since PCache instances are not able to rob
 150 ** unused pages from other PCaches) but it also operates without a mutex,
 151 ** and is therefore often faster.  Mode 2 requires a mutex in order to be
 152 ** threadsafe, but recycles pages more efficiently.
 153 **
 154 ** For mode (1), PGroup.mutex is NULL.  For mode (2) there is only a single
 155 ** PGroup which is the pcache1.grp global variable and its mutex is
 156 ** SQLITE_MUTEX_STATIC_LRU.
 157 */
 158 struct PGroup {
 159   sqlite3_mutex *mutex;          /* MUTEX_STATIC_LRU or NULL */
 160   unsigned int nMaxPage;         /* Sum of nMax for purgeable caches */
 161   unsigned int nMinPage;         /* Sum of nMin for purgeable caches */
 162   unsigned int mxPinned;         /* nMaxpage + 10 - nMinPage */
 163   unsigned int nPurgeable;       /* Number of purgeable pages allocated */
 164   PgHdr1 lru;                    /* The beginning and end of the LRU list */
 165 };
 166
 167 /* Each page cache is an instance of the following object.  Every
 168 ** open database file (including each in-memory database and each
 169 ** temporary or transient database) has a single page cache which
 170 ** is an instance of this object.
 171 **
 172 ** Pointers to structures of this type are cast and returned as
 173 ** opaque sqlite3_pcache* handles.
 174 */
 175 struct PCache1 {
 176   /* Cache configuration parameters. Page size (szPage) and the purgeable
 177   ** flag (bPurgeable) and the pnPurgeable pointer are all set when the
 178   ** cache is created and are never changed thereafter. nMax may be
 179   ** modified at any time by a call to the pcache1Cachesize() method.
 180   ** The PGroup mutex must be held when accessing nMax.
 181   */
 182   PGroup *pGroup;                     /* PGroup this cache belongs to */
 183   unsigned int *pnPurgeable;          /* Pointer to pGroup->nPurgeable */
 184   int szPage;                         /* Size of database content section */
 185   int szExtra;                        /* sizeof(MemPage)+sizeof(PgHdr) */
 186   int szAlloc;                        /* Total size of one pcache line */
 187   int bPurgeable;                     /* True if cache is purgeable */
 188   unsigned int nMin;                  /* Minimum number of pages reserved */
 189   unsigned int nMax;                  /* Configured "cache_size" value */
 190   unsigned int n90pct;                /* nMax*9/10 */
 191   unsigned int iMaxKey;               /* Largest key seen since xTruncate() */
 192   unsigned int nPurgeableDummy;       /* pnPurgeable points here when not used*/
 193
 194   /* Hash table of all pages. The following variables may only be accessed
 195   ** when the accessor is holding the PGroup mutex.
 196   */
 197   unsigned int nRecyclable;           /* Number of pages in the LRU list */
 198   unsigned int nPage;                 /* Total number of pages in apHash */
 199   unsigned int nHash;                 /* Number of slots in apHash[] */
 200   PgHdr1 **apHash;                    /* Hash table for fast lookup by key */
 201   PgHdr1 *pFree;                      /* List of unused pcache-local pages */
 202   void *pBulk;                        /* Bulk memory used by pcache-local */
 203 };
 204
 205 /*
 206 ** Free slots in the allocator used to divide up the global page cache
 207 ** buffer provided using the SQLITE_CONFIG_PAGECACHE mechanism.
 208 */
 209 struct PgFreeslot {
 210   PgFreeslot *pNext;  /* Next free slot */
 211 };
 212
 213 /*
 214 ** Global data used by this cache.
 215 */
 216 static SQLITE_WSD struct PCacheGlobal {
 217   PGroup grp;                    /* The global PGroup for mode (2) */
 218
 219   /* Variables related to SQLITE_CONFIG_PAGECACHE settings.  The
 220   ** szSlot, nSlot, pStart, pEnd, nReserve, and isInit values are all
 221   ** fixed at sqlite3_initialize() time and do not require mutex protection.
 222   ** The nFreeSlot and pFree values do require mutex protection.
 223   */
 224   int isInit;                    /* True if initialized */
 225   int separateCache;             /* Use a new PGroup for each PCache */
 226   int nInitPage;                 /* Initial bulk allocation size */
 227   int szSlot;                    /* Size of each free slot */
 228   int nSlot;                     /* The number of pcache slots */
 229   int nReserve;                  /* Try to keep nFreeSlot above this */
 230   void *pStart, *pEnd;           /* Bounds of global page cache memory */
 231   /* Above requires no mutex.  Use mutex below for variable that follow. */
 232   sqlite3_mutex *mutex;          /* Mutex for accessing the following: */
 233   PgFreeslot *pFree;             /* Free page blocks */
 234   int nFreeSlot;                 /* Number of unused pcache slots */
 235   /* The following value requires a mutex to change.  We skip the mutex on
 236   ** reading because (1) most platforms read a 32-bit integer atomically and
 237   ** (2) even if an incorrect value is read, no great harm is done since this
 238   ** is really just an optimization. */
 239   int bUnderPressure;            /* True if low on PAGECACHE memory */
 240 } pcache1_g;
 241
 242 /*
 243 ** All code in this file should access the global structure above via the
 244 ** alias "pcache1". This ensures that the WSD emulation is used when
 245 ** compiling for systems that do not support real WSD.
 246 */
 247 #define pcache1 (GLOBAL(struct PCacheGlobal, pcache1_g))
 248
 249 /*
 250 ** Macros to enter and leave the PCache LRU mutex.
 251 */
 252 #if !defined(SQLITE_ENABLE_MEMORY_MANAGEMENT) || SQLITE_THREADSAFE==0
 253 # define pcache1EnterMutex(X)  assert((X)->mutex==0)
 254 # define pcache1LeaveMutex(X)  assert((X)->mutex==0)
 255 # define PCACHE1_MIGHT_USE_GROUP_MUTEX 0
 256 #else
 257 # define pcache1EnterMutex(X) sqlite3_mutex_enter((X)->mutex)
 258 # define pcache1LeaveMutex(X) sqlite3_mutex_leave((X)->mutex)
 259 # define PCACHE1_MIGHT_USE_GROUP_MUTEX 1
 260 #endif
 261
 262 /******************************************************************************/
 263 /******** Page Allocation/SQLITE_CONFIG_PCACHE Related Functions **************/
 264
 265
 266 /*
 267 ** This function is called during initialization if a static buffer is
 268 ** supplied to use for the page-cache by passing the SQLITE_CONFIG_PAGECACHE
 269 ** verb to sqlite3_config(). Parameter pBuf points to an allocation large
 270 ** enough to contain 'n' buffers of 'sz' bytes each.
 271 **
 272 ** This routine is called from sqlite3_initialize() and so it is guaranteed
 273 ** to be serialized already.  There is no need for further mutexing.
 274 */
 275 void sqlite3PCacheBufferSetup(void *pBuf, int sz, int n){
 276   if( pcache1.isInit ){
 277     PgFreeslot *p;
 278     if( pBuf==0 ) sz = n = 0;
 279     if( n==0 ) sz = 0;
 280     sz = ROUNDDOWN8(sz);
 281     pcache1.szSlot = sz;
 282     pcache1.nSlot = pcache1.nFreeSlot = n;
 283     pcache1.nReserve = n>90 ? 10 : (n/10 + 1);
 284     pcache1.pStart = pBuf;
 285     pcache1.pFree = 0;
 286     pcache1.bUnderPressure = 0;
 287     while( n-- ){
 288       p = (PgFreeslot*)pBuf;
 289       p->pNext = pcache1.pFree;
 290       pcache1.pFree = p;
 291       pBuf = (void*)&((char*)pBuf)[sz];
 292     }
 293     pcache1.pEnd = pBuf;
 294   }
 295 }
 296
 297 /*
 298 ** Try to initialize the pCache->pFree and pCache->pBulk fields.  Return
 299 ** true if pCache->pFree ends up containing one or more free pages.
 300 */
 301 static int pcache1InitBulk(PCache1 *pCache){
 302   i64 szBulk;
 303   char *zBulk;
 304   if( pcache1.nInitPage==0 ) return 0;
 305   /* Do not bother with a bulk allocation if the cache size very small */
 306   if( pCache->nMax<3 ) return 0;
 307   sqlite3BeginBenignMalloc();
 308   if( pcache1.nInitPage>0 ){
 309     szBulk = pCache->szAlloc * (i64)pcache1.nInitPage;
 310   }else{
 311     szBulk = -1024 * (i64)pcache1.nInitPage;
 312   }
 313   if( szBulk > pCache->szAlloc*(i64)pCache->nMax ){
 314     szBulk = pCache->szAlloc*(i64)pCache->nMax;
 315   }
 316   zBulk = pCache->pBulk = sqlite3Malloc( szBulk );
 317   sqlite3EndBenignMalloc();
 318   if( zBulk ){
 319     int nBulk = sqlite3MallocSize(zBulk)/pCache->szAlloc;
 320     do{
 321       PgHdr1 *pX = (PgHdr1*)&zBulk[pCache->szPage];
 322       pX->page.pBuf = zBulk;
 323       pX->page.pExtra = (u8*)pX + ROUND8(sizeof(*pX));
 324       assert( EIGHT_BYTE_ALIGNMENT( pX->page.pExtra ) );
 325       pX->isBulkLocal = 1;
 326       pX->isAnchor = 0;
 327       pX->pNext = pCache->pFree;
 328       pX->pLruPrev = 0;           /* Initializing this saves a valgrind error */
 329       pCache->pFree = pX;
 330       zBulk += pCache->szAlloc;
 331     }while( --nBulk );
 332   }
 333   return pCache->pFree!=0;
 334 }
 335
 336 /*
 337 ** Malloc function used within this file to allocate space from the buffer
 338 ** configured using sqlite3_config(SQLITE_CONFIG_PAGECACHE) option. If no
 339 ** such buffer exists or there is no space left in it, this function falls
 340 ** back to sqlite3Malloc().
 341 **
 342 ** Multiple threads can run this routine at the same time.  Global variables
 343 ** in pcache1 need to be protected via mutex.
 344 */
 345 static void *pcache1Alloc(int nByte){
 346   void *p = 0;
 347   assert( sqlite3_mutex_notheld(pcache1.grp.mutex) );
 348   if( nByte<=pcache1.szSlot ){
 349     sqlite3_mutex_enter(pcache1.mutex);
 350     p = (PgHdr1 *)pcache1.pFree;
 351     if( p ){
 352       pcache1.pFree = pcache1.pFree->pNext;
 353       pcache1.nFreeSlot--;
 354       pcache1.bUnderPressure = pcache1.nFreeSlot<pcache1.nReserve;
 355       assert( pcache1.nFreeSlot>=0 );
 356       sqlite3StatusHighwater(SQLITE_STATUS_PAGECACHE_SIZE, nByte);
 357       sqlite3StatusUp(SQLITE_STATUS_PAGECACHE_USED, 1);
 358     }
 359     sqlite3_mutex_leave(pcache1.mutex);
 360   }
 361   if( p==0 ){
 362     /* Memory is not available in the SQLITE_CONFIG_PAGECACHE pool.  Get
 363     ** it from sqlite3Malloc instead.
 364     */
 365     p = sqlite3Malloc(nByte);
 366 #ifndef SQLITE_DISABLE_PAGECACHE_OVERFLOW_STATS
 367     if( p ){
 368       int sz = sqlite3MallocSize(p);
 369       sqlite3_mutex_enter(pcache1.mutex);
 370       sqlite3StatusHighwater(SQLITE_STATUS_PAGECACHE_SIZE, nByte);
 371       sqlite3StatusUp(SQLITE_STATUS_PAGECACHE_OVERFLOW, sz);
 372       sqlite3_mutex_leave(pcache1.mutex);
 373     }
 374 #endif
 375     sqlite3MemdebugSetType(p, MEMTYPE_PCACHE);
 376   }
 377   return p;
 378 }
 379
 380 /*
 381 ** Free an allocated buffer obtained from pcache1Alloc().
 382 */
 383 static void pcache1Free(void *p){
 384   if( p==0 ) return;
 385   if( SQLITE_WITHIN(p, pcache1.pStart, pcache1.pEnd) ){
 386     PgFreeslot *pSlot;
 387     sqlite3_mutex_enter(pcache1.mutex);
 388     sqlite3StatusDown(SQLITE_STATUS_PAGECACHE_USED, 1);
 389     pSlot = (PgFreeslot*)p;
 390     pSlot->pNext = pcache1.pFree;
 391     pcache1.pFree = pSlot;
 392     pcache1.nFreeSlot++;
 393     pcache1.bUnderPressure = pcache1.nFreeSlot<pcache1.nReserve;
 394     assert( pcache1.nFreeSlot<=pcache1.nSlot );
 395     sqlite3_mutex_leave(pcache1.mutex);
 396   }else{
 397     assert( sqlite3MemdebugHasType(p, MEMTYPE_PCACHE) );
 398     sqlite3MemdebugSetType(p, MEMTYPE_HEAP);
 399 #ifndef SQLITE_DISABLE_PAGECACHE_OVERFLOW_STATS
 400     {
 401       int nFreed = 0;
 402       nFreed = sqlite3MallocSize(p);
 403       sqlite3_mutex_enter(pcache1.mutex);
 404       sqlite3StatusDown(SQLITE_STATUS_PAGECACHE_OVERFLOW, nFreed);
 405       sqlite3_mutex_leave(pcache1.mutex);
 406     }
 407 #endif
 408     sqlite3_free(p);
 409   }
 410 }
 411
 412 #ifdef SQLITE_ENABLE_MEMORY_MANAGEMENT
 413 /*
 414 ** Return the size of a pcache allocation
 415 */
 416 static int pcache1MemSize(void *p){
 417   if( p>=pcache1.pStart && p<pcache1.pEnd ){
 418     return pcache1.szSlot;
 419   }else{
 420     int iSize;
 421     assert( sqlite3MemdebugHasType(p, MEMTYPE_PCACHE) );
 422     sqlite3MemdebugSetType(p, MEMTYPE_HEAP);
 423     iSize = sqlite3MallocSize(p);
 424     sqlite3MemdebugSetType(p, MEMTYPE_PCACHE);
 425     return iSize;
 426   }
 427 }
 428 #endif /* SQLITE_ENABLE_MEMORY_MANAGEMENT */
 429
 430 /*
 431 ** Allocate a new page object initially associated with cache pCache.
 432 */
 433 static PgHdr1 *pcache1AllocPage(PCache1 *pCache, int benignMalloc){
 434   PgHdr1 *p = 0;
 435   void *pPg;
 436
 437   assert( sqlite3_mutex_held(pCache->pGroup->mutex) );
 438   if( pCache->pFree || (pCache->nPage==0 && pcache1InitBulk(pCache)) ){
 439     assert( pCache->pFree!=0 );
 440     p = pCache->pFree;
 441     pCache->pFree = p->pNext;
 442     p->pNext = 0;
 443   }else{
 444 #ifdef SQLITE_ENABLE_MEMORY_MANAGEMENT
 445     /* The group mutex must be released before pcache1Alloc() is called. This
 446     ** is because it might call sqlite3_release_memory(), which assumes that
 447     ** this mutex is not held. */
 448     assert( pcache1.separateCache==0 );
 449     assert( pCache->pGroup==&pcache1.grp );
 450     pcache1LeaveMutex(pCache->pGroup);
 451 #endif
 452     if( benignMalloc ){ sqlite3BeginBenignMalloc(); }
 453     pPg = pcache1Alloc(pCache->szAlloc);
 454     if( benignMalloc ){ sqlite3EndBenignMalloc(); }
 455 #ifdef SQLITE_ENABLE_MEMORY_MANAGEMENT
 456     pcache1EnterMutex(pCache->pGroup);
 457 #endif
 458     if( pPg==0 ) return 0;
 459     p = (PgHdr1 *)&((u8 *)pPg)[pCache->szPage];
 460     p->page.pBuf = pPg;
 461     p->page.pExtra = (u8*)p + ROUND8(sizeof(*p));
 462     assert( EIGHT_BYTE_ALIGNMENT( p->page.pExtra ) );
 463     p->isBulkLocal = 0;
 464     p->isAnchor = 0;
 465     p->pLruPrev = 0;           /* Initializing this saves a valgrind error */
 466   }
 467   (*pCache->pnPurgeable)++;
 468   return p;
 469 }
 470
 471 /*
 472 ** Free a page object allocated by pcache1AllocPage().
 473 */
 474 static void pcache1FreePage(PgHdr1 *p){
 475   PCache1 *pCache;
 476   assert( p!=0 );
 477   pCache = p->pCache;
 478   assert( sqlite3_mutex_held(p->pCache->pGroup->mutex) );
 479   if( p->isBulkLocal ){
 480     p->pNext = pCache->pFree;
 481     pCache->pFree = p;
 482   }else{
 483     pcache1Free(p->page.pBuf);
 484   }
 485   (*pCache->pnPurgeable)--;
 486 }
 487
 488 /*
 489 ** Malloc function used by SQLite to obtain space from the buffer configured
 490 ** using sqlite3_config(SQLITE_CONFIG_PAGECACHE) option. If no such buffer
 491 ** exists, this function falls back to sqlite3Malloc().
 492 */
 493 void *sqlite3PageMalloc(int sz){
 494   assert( sz<=65536+8 ); /* These allocations are never very large */
 495   return pcache1Alloc(sz);
 496 }
 497
 498 /*
 499 ** Free an allocated buffer obtained from sqlite3PageMalloc().
 500 */
 501 void sqlite3PageFree(void *p){
 502   pcache1Free(p);
 503 }
 504
 505
 506 /*
 507 ** Return true if it desirable to avoid allocating a new page cache
 508 ** entry.
 509 **
 510 ** If memory was allocated specifically to the page cache using
 511 ** SQLITE_CONFIG_PAGECACHE but that memory has all been used, then
 512 ** it is desirable to avoid allocating a new page cache entry because
 513 ** presumably SQLITE_CONFIG_PAGECACHE was suppose to be sufficient
 514 ** for all page cache needs and we should not need to spill the
 515 ** allocation onto the heap.
 516 **
 517 ** Or, the heap is used for all page cache memory but the heap is
 518 ** under memory pressure, then again it is desirable to avoid
 519 ** allocating a new page cache entry in order to avoid stressing
 520 ** the heap even further.
 521 */
 522 static int pcache1UnderMemoryPressure(PCache1 *pCache){
 523   if( pcache1.nSlot && (pCache->szPage+pCache->szExtra)<=pcache1.szSlot ){
 524     return pcache1.bUnderPressure;
 525   }else{
 526     return sqlite3HeapNearlyFull();
 527   }
 528 }
 529
 530 /******************************************************************************/
 531 /******** General Implementation Functions ************************************/
 532
 533 /*
 534 ** This function is used to resize the hash table used by the cache passed
 535 ** as the first argument.
 536 **
 537 ** The PCache mutex must be held when this function is called.
 538 */
 539 static void pcache1ResizeHash(PCache1 *p){
 540   PgHdr1 **apNew;
 541   unsigned int nNew;
 542   unsigned int i;
 543
 544   assert( sqlite3_mutex_held(p->pGroup->mutex) );
 545
 546   nNew = p->nHash*2;
 547   if( nNew<256 ){
 548     nNew = 256;
 549   }
 550
 551   pcache1LeaveMutex(p->pGroup);
 552   if( p->nHash ){ sqlite3BeginBenignMalloc(); }
 553   apNew = (PgHdr1 **)sqlite3MallocZero(sizeof(PgHdr1 *)*nNew);
 554   if( p->nHash ){ sqlite3EndBenignMalloc(); }
 555   pcache1EnterMutex(p->pGroup);
 556   if( apNew ){
 557     for(i=0; i<p->nHash; i++){
 558       PgHdr1 *pPage;
 559       PgHdr1 *pNext = p->apHash[i];
 560       while( (pPage = pNext)!=0 ){
 561         unsigned int h = pPage->iKey % nNew;
 562         pNext = pPage->pNext;
 563         pPage->pNext = apNew[h];
 564         apNew[h] = pPage;
 565       }
 566     }
 567     sqlite3_free(p->apHash);
 568     p->apHash = apNew;
 569     p->nHash = nNew;
 570   }
 571 }
 572
 573 /*
 574 ** This function is used internally to remove the page pPage from the
 575 ** PGroup LRU list, if is part of it. If pPage is not part of the PGroup
 576 ** LRU list, then this function is a no-op.
 577 **
 578 ** The PGroup mutex must be held when this function is called.
 579 */
 580 static PgHdr1 *pcache1PinPage(PgHdr1 *pPage){
 581   assert( pPage!=0 );
 582   assert( PAGE_IS_UNPINNED(pPage) );
 583   assert( pPage->pLruNext );
 584   assert( pPage->pLruPrev );
 585   assert( sqlite3_mutex_held(pPage->pCache->pGroup->mutex) );
 586   pPage->pLruPrev->pLruNext = pPage->pLruNext;
 587   pPage->pLruNext->pLruPrev = pPage->pLruPrev;
 588   pPage->pLruNext = 0;
 589   /* pPage->pLruPrev = 0;
 590   ** No need to clear pLruPrev as it is never accessed if pLruNext is 0 */
 591   assert( pPage->isAnchor==0 );
 592   assert( pPage->pCache->pGroup->lru.isAnchor==1 );
 593   pPage->pCache->nRecyclable--;
 594   return pPage;
 595 }
 596
 597
 598 /*
 599 ** Remove the page supplied as an argument from the hash table
 600 ** (PCache1.apHash structure) that it is currently stored in.
 601 ** Also free the page if freePage is true.
 602 **
 603 ** The PGroup mutex must be held when this function is called.
 604 */
 605 static void pcache1RemoveFromHash(PgHdr1 *pPage, int freeFlag){
 606   unsigned int h;
 607   PCache1 *pCache = pPage->pCache;
 608   PgHdr1 **pp;
 609
 610   assert( sqlite3_mutex_held(pCache->pGroup->mutex) );
 611   h = pPage->iKey % pCache->nHash;
 612   for(pp=&pCache->apHash[h]; (*pp)!=pPage; pp=&(*pp)->pNext);
 613   *pp = (*pp)->pNext;
 614
 615   pCache->nPage--;
 616   if( freeFlag ) pcache1FreePage(pPage);
 617 }
 618
 619 /*
 620 ** If there are currently more than nMaxPage pages allocated, try
 621 ** to recycle pages to reduce the number allocated to nMaxPage.
 622 */
 623 static void pcache1EnforceMaxPage(PCache1 *pCache){
 624   PGroup *pGroup = pCache->pGroup;
 625   PgHdr1 *p;
 626   assert( sqlite3_mutex_held(pGroup->mutex) );
 627   while( pGroup->nPurgeable>pGroup->nMaxPage
 628       && (p=pGroup->lru.pLruPrev)->isAnchor==0
 629   ){
 630     assert( p->pCache->pGroup==pGroup );
 631     assert( PAGE_IS_UNPINNED(p) );
 632     pcache1PinPage(p);
 633     pcache1RemoveFromHash(p, 1);
 634   }
 635   if( pCache->nPage==0 && pCache->pBulk ){
 636     sqlite3_free(pCache->pBulk);
 637     pCache->pBulk = pCache->pFree = 0;
 638   }
 639 }
 640
 641 /*
 642 ** Discard all pages from cache pCache with a page number (key value)
 643 ** greater than or equal to iLimit. Any pinned pages that meet this
 644 ** criteria are unpinned before they are discarded.
 645 **
 646 ** The PCache mutex must be held when this function is called.
 647 */
 648 static void pcache1TruncateUnsafe(
 649   PCache1 *pCache,             /* The cache to truncate */
 650   unsigned int iLimit          /* Drop pages with this pgno or larger */
 651 ){
 652   TESTONLY( int nPage = 0; )  /* To assert pCache->nPage is correct */
 653   unsigned int h, iStop;
 654   assert( sqlite3_mutex_held(pCache->pGroup->mutex) );
 655   assert( pCache->iMaxKey >= iLimit );
 656   assert( pCache->nHash > 0 );
 657   if( pCache->iMaxKey - iLimit < pCache->nHash ){
 658     /* If we are just shaving the last few pages off the end of the
 659     ** cache, then there is no point in scanning the entire hash table.
 660     ** Only scan those hash slots that might contain pages that need to
 661     ** be removed. */
 662     h = iLimit % pCache->nHash;
 663     iStop = pCache->iMaxKey % pCache->nHash;
 664     TESTONLY( nPage = -10; )  /* Disable the pCache->nPage validity check */
 665   }else{
 666     /* This is the general case where many pages are being removed.
 667     ** It is necessary to scan the entire hash table */
 668     h = pCache->nHash/2;
 669     iStop = h - 1;
 670   }
 671   for(;;){
 672     PgHdr1 **pp;
 673     PgHdr1 *pPage;
 674     assert( h<pCache->nHash );
 675     pp = &pCache->apHash[h];
 676     while( (pPage = *pp)!=0 ){
 677       if( pPage->iKey>=iLimit ){
 678         pCache->nPage--;
 679         *pp = pPage->pNext;
 680         if( PAGE_IS_UNPINNED(pPage) ) pcache1PinPage(pPage);
 681         pcache1FreePage(pPage);
 682       }else{
 683         pp = &pPage->pNext;
 684         TESTONLY( if( nPage>=0 ) nPage++; )
 685       }
 686     }
 687     if( h==iStop ) break;
 688     h = (h+1) % pCache->nHash;
 689   }
 690   assert( nPage<0 || pCache->nPage==(unsigned)nPage );
 691 }
 692
 693 /******************************************************************************/
 694 /******** sqlite3_pcache Methods **********************************************/
 695
 696 /*
 697 ** Implementation of the sqlite3_pcache.xInit method.
 698 */
 699 static int pcache1Init(void *NotUsed){
 700   UNUSED_PARAMETER(NotUsed);
 701   assert( pcache1.isInit==0 );
 702   memset(&pcache1, 0, sizeof(pcache1));
 703
 704
 705   /*
 706   ** The pcache1.separateCache variable is true if each PCache has its own
 707   ** private PGroup (mode-1).  pcache1.separateCache is false if the single
 708   ** PGroup in pcache1.grp is used for all page caches (mode-2).
 709   **
 710   **   *  Always use a unified cache (mode-2) if ENABLE_MEMORY_MANAGEMENT
 711   **
 712   **   *  Use a unified cache in single-threaded applications that have
 713   **      configured a start-time buffer for use as page-cache memory using
 714   **      sqlite3_config(SQLITE_CONFIG_PAGECACHE, pBuf, sz, N) with non-NULL
 715   **      pBuf argument.
 716   **
 717   **   *  Otherwise use separate caches (mode-1)
 718   */
 719 #if defined(SQLITE_ENABLE_MEMORY_MANAGEMENT)
 720   pcache1.separateCache = 0;
 721 #elif SQLITE_THREADSAFE
 722   pcache1.separateCache = sqlite3GlobalConfig.pPage==0
 723                           || sqlite3GlobalConfig.bCoreMutex>0;
 724 #else
 725   pcache1.separateCache = sqlite3GlobalConfig.pPage==0;
 726 #endif
 727
 728 #if SQLITE_THREADSAFE
 729   if( sqlite3GlobalConfig.bCoreMutex ){
 730     pcache1.grp.mutex = sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_LRU);
 731     pcache1.mutex = sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_PMEM);
 732   }
 733 #endif
 734   if( pcache1.separateCache
 735    && sqlite3GlobalConfig.nPage!=0
 736    && sqlite3GlobalConfig.pPage==0
 737   ){
 738     pcache1.nInitPage = sqlite3GlobalConfig.nPage;
 739   }else{
 740     pcache1.nInitPage = 0;
 741   }
 742   pcache1.grp.mxPinned = 10;
 743   pcache1.isInit = 1;
 744   return SQLITE_OK;
 745 }
 746
 747 /*
 748 ** Implementation of the sqlite3_pcache.xShutdown method.
 749 ** Note that the static mutex allocated in xInit does
 750 ** not need to be freed.
 751 */
 752 static void pcache1Shutdown(void *NotUsed){
 753   UNUSED_PARAMETER(NotUsed);
 754   assert( pcache1.isInit!=0 );
 755   memset(&pcache1, 0, sizeof(pcache1));
 756 }
 757
 758 /* forward declaration */
 759 static void pcache1Destroy(sqlite3_pcache *p);
 760
 761 /*
 762 ** Implementation of the sqlite3_pcache.xCreate method.
 763 **
 764 ** Allocate a new cache.
 765 */
 766 static sqlite3_pcache *pcache1Create(int szPage, int szExtra, int bPurgeable){
 767   PCache1 *pCache;      /* The newly created page cache */
 768   PGroup *pGroup;       /* The group the new page cache will belong to */
 769   int sz;               /* Bytes of memory required to allocate the new cache */
 770
 771   assert( (szPage & (szPage-1))==0 && szPage>=512 && szPage<=65536 );
 772   assert( szExtra < 300 );
 773
 774   sz = sizeof(PCache1) + sizeof(PGroup)*pcache1.separateCache;
 775   pCache = (PCache1 *)sqlite3MallocZero(sz);
 776   if( pCache ){
 777     if( pcache1.separateCache ){
 778       pGroup = (PGroup*)&pCache[1];
 779       pGroup->mxPinned = 10;
 780     }else{
 781       pGroup = &pcache1.grp;
 782     }
 783     pcache1EnterMutex(pGroup);
 784     if( pGroup->lru.isAnchor==0 ){
 785       pGroup->lru.isAnchor = 1;
 786       pGroup->lru.pLruPrev = pGroup->lru.pLruNext = &pGroup->lru;
 787     }
 788     pCache->pGroup = pGroup;
 789     pCache->szPage = szPage;
 790     pCache->szExtra = szExtra;
 791     pCache->szAlloc = szPage + szExtra + ROUND8(sizeof(PgHdr1));
 792     pCache->bPurgeable = (bPurgeable ? 1 : 0);
 793     pcache1ResizeHash(pCache);
 794     if( bPurgeable ){
 795       pCache->nMin = 10;
 796       pGroup->nMinPage += pCache->nMin;
 797       pGroup->mxPinned = pGroup->nMaxPage + 10 - pGroup->nMinPage;
 798       pCache->pnPurgeable = &pGroup->nPurgeable;
 799     }else{
 800       pCache->pnPurgeable = &pCache->nPurgeableDummy;
 801     }
 802     pcache1LeaveMutex(pGroup);
 803     if( pCache->nHash==0 ){
 804       pcache1Destroy((sqlite3_pcache*)pCache);
 805       pCache = 0;
 806     }
 807   }
 808   return (sqlite3_pcache *)pCache;
 809 }
 810
 811 /*
 812 ** Implementation of the sqlite3_pcache.xCachesize method.
 813 **
 814 ** Configure the cache_size limit for a cache.
 815 */
 816 static void pcache1Cachesize(sqlite3_pcache *p, int nMax){
 817   PCache1 *pCache = (PCache1 *)p;
 818   u32 n;
 819   assert( nMax>=0 );
 820   if( pCache->bPurgeable ){
 821     PGroup *pGroup = pCache->pGroup;
 822     pcache1EnterMutex(pGroup);
 823     n = (u32)nMax;
 824     if( n > 0x7fff0000 - pGroup->nMaxPage + pCache->nMax ){
 825       n = 0x7fff0000 - pGroup->nMaxPage + pCache->nMax;
 826     }
 827     pGroup->nMaxPage += (n - pCache->nMax);
 828     pGroup->mxPinned = pGroup->nMaxPage + 10 - pGroup->nMinPage;
 829     pCache->nMax = n;
 830     pCache->n90pct = pCache->nMax*9/10;
 831     pcache1EnforceMaxPage(pCache);
 832     pcache1LeaveMutex(pGroup);
 833   }
 834 }
 835
 836 /*
 837 ** Implementation of the sqlite3_pcache.xShrink method.
 838 **
 839 ** Free up as much memory as possible.
 840 */
 841 static void pcache1Shrink(sqlite3_pcache *p){
 842   PCache1 *pCache = (PCache1*)p;
 843   if( pCache->bPurgeable ){
 844     PGroup *pGroup = pCache->pGroup;
 845     unsigned int savedMaxPage;
 846     pcache1EnterMutex(pGroup);
 847     savedMaxPage = pGroup->nMaxPage;
 848     pGroup->nMaxPage = 0;
 849     pcache1EnforceMaxPage(pCache);
 850     pGroup->nMaxPage = savedMaxPage;
 851     pcache1LeaveMutex(pGroup);
 852   }
 853 }
 854
 855 /*
 856 ** Implementation of the sqlite3_pcache.xPagecount method.
 857 */
 858 static int pcache1Pagecount(sqlite3_pcache *p){
 859   int n;
 860   PCache1 *pCache = (PCache1*)p;
 861   pcache1EnterMutex(pCache->pGroup);
 862   n = pCache->nPage;
 863   pcache1LeaveMutex(pCache->pGroup);
 864   return n;
 865 }
 866
 867
 868 /*
 869 ** Implement steps 3, 4, and 5 of the pcache1Fetch() algorithm described
 870 ** in the header of the pcache1Fetch() procedure.
 871 **
 872 ** This steps are broken out into a separate procedure because they are
 873 ** usually not needed, and by avoiding the stack initialization required
 874 ** for these steps, the main pcache1Fetch() procedure can run faster.
 875 */
 876 static SQLITE_NOINLINE PgHdr1 *pcache1FetchStage2(
 877   PCache1 *pCache,
 878   unsigned int iKey,
 879   int createFlag
 880 ){
 881   unsigned int nPinned;
 882   PGroup *pGroup = pCache->pGroup;
 883   PgHdr1 *pPage = 0;
 884
 885   /* Step 3: Abort if createFlag is 1 but the cache is nearly full */
 886   assert( pCache->nPage >= pCache->nRecyclable );
 887   nPinned = pCache->nPage - pCache->nRecyclable;
 888   assert( pGroup->mxPinned == pGroup->nMaxPage + 10 - pGroup->nMinPage );
 889   assert( pCache->n90pct == pCache->nMax*9/10 );
 890   if( createFlag==1 && (
 891         nPinned>=pGroup->mxPinned
 892      || nPinned>=pCache->n90pct
 893      || (pcache1UnderMemoryPressure(pCache) && pCache->nRecyclable<nPinned)
 894   )){
 895     return 0;
 896   }
 897
 898   if( pCache->nPage>=pCache->nHash ) pcache1ResizeHash(pCache);
 899   assert( pCache->nHash>0 && pCache->apHash );
 900
 901   /* Step 4. Try to recycle a page. */
 902   if( pCache->bPurgeable
 903    && !pGroup->lru.pLruPrev->isAnchor
 904    && ((pCache->nPage+1>=pCache->nMax) || pcache1UnderMemoryPressure(pCache))
 905   ){
 906     PCache1 *pOther;
 907     pPage = pGroup->lru.pLruPrev;
 908     assert( PAGE_IS_UNPINNED(pPage) );
 909     pcache1RemoveFromHash(pPage, 0);
 910     pcache1PinPage(pPage);
 911     pOther = pPage->pCache;
 912     if( pOther->szAlloc != pCache->szAlloc ){
 913       pcache1FreePage(pPage);
 914       pPage = 0;
 915     }else{
 916       pGroup->nPurgeable -= (pOther->bPurgeable - pCache->bPurgeable);
 917     }
 918   }
 919
 920   /* Step 5. If a usable page buffer has still not been found,
 921   ** attempt to allocate a new one.
 922   */
 923   if( !pPage ){
 924     pPage = pcache1AllocPage(pCache, createFlag==1);
 925   }
 926
 927   if( pPage ){
 928     unsigned int h = iKey % pCache->nHash;
 929     pCache->nPage++;
 930     pPage->iKey = iKey;
 931     pPage->pNext = pCache->apHash[h];
 932     pPage->pCache = pCache;
 933     pPage->pLruNext = 0;
 934     /* pPage->pLruPrev = 0;
 935     ** No need to clear pLruPrev since it is not accessed when pLruNext==0 */
 936     *(void **)pPage->page.pExtra = 0;
 937     pCache->apHash[h] = pPage;
 938     if( iKey>pCache->iMaxKey ){
 939       pCache->iMaxKey = iKey;
 940     }
 941   }
 942   return pPage;
 943 }
 944
 945 /*
 946 ** Implementation of the sqlite3_pcache.xFetch method.
 947 **
 948 ** Fetch a page by key value.
 949 **
 950 ** Whether or not a new page may be allocated by this function depends on
 951 ** the value of the createFlag argument.  0 means do not allocate a new
 952 ** page.  1 means allocate a new page if space is easily available.  2
 953 ** means to try really hard to allocate a new page.
 954 **
 955 ** For a non-purgeable cache (a cache used as the storage for an in-memory
 956 ** database) there is really no difference between createFlag 1 and 2.  So
 957 ** the calling function (pcache.c) will never have a createFlag of 1 on
 958 ** a non-purgeable cache.
 959 **
 960 ** There are three different approaches to obtaining space for a page,
 961 ** depending on the value of parameter createFlag (which may be 0, 1 or 2).
 962 **
 963 **   1. Regardless of the value of createFlag, the cache is searched for a
 964 **      copy of the requested page. If one is found, it is returned.
 965 **
 966 **   2. If createFlag==0 and the page is not already in the cache, NULL is
 967 **      returned.
 968 **
 969 **   3. If createFlag is 1, and the page is not already in the cache, then
 970 **      return NULL (do not allocate a new page) if any of the following
 971 **      conditions are true:
 972 **
 973 **       (a) the number of pages pinned by the cache is greater than
 974 **           PCache1.nMax, or
 975 **
 976 **       (b) the number of pages pinned by the cache is greater than
 977 **           the sum of nMax for all purgeable caches, less the sum of
 978 **           nMin for all other purgeable caches, or
 979 **
 980 **   4. If none of the first three conditions apply and the cache is marked
 981 **      as purgeable, and if one of the following is true:
 982 **
 983 **       (a) The number of pages allocated for the cache is already
 984 **           PCache1.nMax, or
 985 **
 986 **       (b) The number of pages allocated for all purgeable caches is
 987 **           already equal to or greater than the sum of nMax for all
 988 **           purgeable caches,
 989 **
 990 **       (c) The system is under memory pressure and wants to avoid
 991 **           unnecessary pages cache entry allocations
 992 **
 993 **      then attempt to recycle a page from the LRU list. If it is the right
 994 **      size, return the recycled buffer. Otherwise, free the buffer and
 995 **      proceed to step 5.
 996 **
 997 **   5. Otherwise, allocate and return a new page buffer.
 998 **
 999 ** There are two versions of this routine.  pcache1FetchWithMutex() is
1000 ** the general case.  pcache1FetchNoMutex() is a faster implementation for
1001 ** the common case where pGroup->mutex is NULL.  The pcache1Fetch() wrapper
1002 ** invokes the appropriate routine.
1003 */
1004 static PgHdr1 *pcache1FetchNoMutex(
1005   sqlite3_pcache *p,
1006   unsigned int iKey,
1007   int createFlag
1008 ){
1009   PCache1 *pCache = (PCache1 *)p;
1010   PgHdr1 *pPage = 0;
1011
1012   /* Step 1: Search the hash table for an existing entry. */
1013   pPage = pCache->apHash[iKey % pCache->nHash];
1014   while( pPage && pPage->iKey!=iKey ){ pPage = pPage->pNext; }
1015
1016   /* Step 2: If the page was found in the hash table, then return it.
1017   ** If the page was not in the hash table and createFlag is 0, abort.
1018   ** Otherwise (page not in hash and createFlag!=0) continue with
1019   ** subsequent steps to try to create the page. */
1020   if( pPage ){
1021     if( PAGE_IS_UNPINNED(pPage) ){
1022       return pcache1PinPage(pPage);
1023     }else{
1024       return pPage;
1025     }
1026   }else if( createFlag ){
1027     /* Steps 3, 4, and 5 implemented by this subroutine */
1028     return pcache1FetchStage2(pCache, iKey, createFlag);
1029   }else{
1030     return 0;
1031   }
1032 }
1033 #if PCACHE1_MIGHT_USE_GROUP_MUTEX
1034 static PgHdr1 *pcache1FetchWithMutex(
1035   sqlite3_pcache *p,
1036   unsigned int iKey,
1037   int createFlag
1038 ){
1039   PCache1 *pCache = (PCache1 *)p;
1040   PgHdr1 *pPage;
1041
1042   pcache1EnterMutex(pCache->pGroup);
1043   pPage = pcache1FetchNoMutex(p, iKey, createFlag);
1044   assert( pPage==0 || pCache->iMaxKey>=iKey );
1045   pcache1LeaveMutex(pCache->pGroup);
1046   return pPage;
1047 }
1048 #endif
1049 static sqlite3_pcache_page *pcache1Fetch(
1050   sqlite3_pcache *p,
1051   unsigned int iKey,
1052   int createFlag
1053 ){
1054 #if PCACHE1_MIGHT_USE_GROUP_MUTEX || defined(SQLITE_DEBUG)
1055   PCache1 *pCache = (PCache1 *)p;
1056 #endif
1057
1058   assert( offsetof(PgHdr1,page)==0 );
1059   assert( pCache->bPurgeable || createFlag!=1 );
1060   assert( pCache->bPurgeable || pCache->nMin==0 );
1061   assert( pCache->bPurgeable==0 || pCache->nMin==10 );
1062   assert( pCache->nMin==0 || pCache->bPurgeable );
1063   assert( pCache->nHash>0 );
1064 #if PCACHE1_MIGHT_USE_GROUP_MUTEX
1065   if( pCache->pGroup->mutex ){
1066     return (sqlite3_pcache_page*)pcache1FetchWithMutex(p, iKey, createFlag);
1067   }else
1068 #endif
1069   {
1070     return (sqlite3_pcache_page*)pcache1FetchNoMutex(p, iKey, createFlag);
1071   }
1072 }
1073
1074
1075 /*
1076 ** Implementation of the sqlite3_pcache.xUnpin method.
1077 **
1078 ** Mark a page as unpinned (eligible for asynchronous recycling).
1079 */
1080 static void pcache1Unpin(
1081   sqlite3_pcache *p,
1082   sqlite3_pcache_page *pPg,
1083   int reuseUnlikely
1084 ){
1085   PCache1 *pCache = (PCache1 *)p;
1086   PgHdr1 *pPage = (PgHdr1 *)pPg;
1087   PGroup *pGroup = pCache->pGroup;
1088
1089   assert( pPage->pCache==pCache );
1090   pcache1EnterMutex(pGroup);
1091
1092   /* It is an error to call this function if the page is already
1093   ** part of the PGroup LRU list.
1094   */
1095   assert( pPage->pLruNext==0 );
1096   assert( PAGE_IS_PINNED(pPage) );
1097
1098   if( reuseUnlikely || pGroup->nPurgeable>pGroup->nMaxPage ){
1099     pcache1RemoveFromHash(pPage, 1);
1100   }else{
1101     /* Add the page to the PGroup LRU list. */
1102     PgHdr1 **ppFirst = &pGroup->lru.pLruNext;
1103     pPage->pLruPrev = &pGroup->lru;
1104     (pPage->pLruNext = *ppFirst)->pLruPrev = pPage;
1105     *ppFirst = pPage;
1106     pCache->nRecyclable++;
1107   }
1108
1109   pcache1LeaveMutex(pCache->pGroup);
1110 }
1111
1112 /*
1113 ** Implementation of the sqlite3_pcache.xRekey method.
1114 */
1115 static void pcache1Rekey(
1116   sqlite3_pcache *p,
1117   sqlite3_pcache_page *pPg,
1118   unsigned int iOld,
1119   unsigned int iNew
1120 ){
1121   PCache1 *pCache = (PCache1 *)p;
1122   PgHdr1 *pPage = (PgHdr1 *)pPg;
1123   PgHdr1 **pp;
1124   unsigned int hOld, hNew;
1125   assert( pPage->iKey==iOld );
1126   assert( pPage->pCache==pCache );
1127   assert( iOld!=iNew );               /* The page number really is changing */
1128
1129   pcache1EnterMutex(pCache->pGroup);
1130
1131   assert( pcache1FetchNoMutex(p, iOld, 0)==pPage ); /* pPg really is iOld */
1132   hOld = iOld%pCache->nHash;
1133   pp = &pCache->apHash[hOld];
1134   while( (*pp)!=pPage ){
1135     pp = &(*pp)->pNext;
1136   }
1137   *pp = pPage->pNext;
1138
1139   assert( pcache1FetchNoMutex(p, iNew, 0)==0 ); /* iNew not in cache */
1140   hNew = iNew%pCache->nHash;
1141   pPage->iKey = iNew;
1142   pPage->pNext = pCache->apHash[hNew];
1143   pCache->apHash[hNew] = pPage;
1144   if( iNew>pCache->iMaxKey ){
1145     pCache->iMaxKey = iNew;
1146   }
1147
1148   pcache1LeaveMutex(pCache->pGroup);
1149 }
1150
1151 /*
1152 ** Implementation of the sqlite3_pcache.xTruncate method.
1153 **
1154 ** Discard all unpinned pages in the cache with a page number equal to
1155 ** or greater than parameter iLimit. Any pinned pages with a page number
1156 ** equal to or greater than iLimit are implicitly unpinned.
1157 */
1158 static void pcache1Truncate(sqlite3_pcache *p, unsigned int iLimit){
1159   PCache1 *pCache = (PCache1 *)p;
1160   pcache1EnterMutex(pCache->pGroup);
1161   if( iLimit<=pCache->iMaxKey ){
1162     pcache1TruncateUnsafe(pCache, iLimit);
1163     pCache->iMaxKey = iLimit-1;
1164   }
1165   pcache1LeaveMutex(pCache->pGroup);
1166 }
1167
1168 /*
1169 ** Implementation of the sqlite3_pcache.xDestroy method.
1170 **
1171 ** Destroy a cache allocated using pcache1Create().
1172 */
1173 static void pcache1Destroy(sqlite3_pcache *p){
1174   PCache1 *pCache = (PCache1 *)p;
1175   PGroup *pGroup = pCache->pGroup;
1176   assert( pCache->bPurgeable || (pCache->nMax==0 && pCache->nMin==0) );
1177   pcache1EnterMutex(pGroup);
1178   if( pCache->nPage ) pcache1TruncateUnsafe(pCache, 0);
1179   assert( pGroup->nMaxPage >= pCache->nMax );
1180   pGroup->nMaxPage -= pCache->nMax;
1181   assert( pGroup->nMinPage >= pCache->nMin );
1182   pGroup->nMinPage -= pCache->nMin;
1183   pGroup->mxPinned = pGroup->nMaxPage + 10 - pGroup->nMinPage;
1184   pcache1EnforceMaxPage(pCache);
1185   pcache1LeaveMutex(pGroup);
1186   sqlite3_free(pCache->pBulk);
1187   sqlite3_free(pCache->apHash);
1188   sqlite3_free(pCache);
1189 }
1190
1191 /*
1192 ** This function is called during initialization (sqlite3_initialize()) to
1193 ** install the default pluggable cache module, assuming the user has not
1194 ** already provided an alternative.
1195 */
1196 void sqlite3PCacheSetDefault(void){
1197   static const sqlite3_pcache_methods2 defaultMethods = {
1198     1,                       /* iVersion */
1199     0,                       /* pArg */
1200     pcache1Init,             /* xInit */
1201     pcache1Shutdown,         /* xShutdown */
1202     pcache1Create,           /* xCreate */
1203     pcache1Cachesize,        /* xCachesize */
1204     pcache1Pagecount,        /* xPagecount */
1205     pcache1Fetch,            /* xFetch */
1206     pcache1Unpin,            /* xUnpin */
1207     pcache1Rekey,            /* xRekey */
1208     pcache1Truncate,         /* xTruncate */
1209     pcache1Destroy,          /* xDestroy */
1210     pcache1Shrink            /* xShrink */
1211   };
1212   sqlite3_config(SQLITE_CONFIG_PCACHE2, &defaultMethods);
1213 }
1214
1215 /*
1216 ** Return the size of the header on each page of this PCACHE implementation.
1217 */
1218 int sqlite3HeaderSizePcache1(void){ return ROUND8(sizeof(PgHdr1)); }
1219
1220 /*
1221 ** Return the global mutex used by this PCACHE implementation.  The
1222 ** sqlite3_status() routine needs access to this mutex.
1223 */
1224 sqlite3_mutex *sqlite3Pcache1Mutex(void){
1225   return pcache1.mutex;
1226 }
1227
1228 #ifdef SQLITE_ENABLE_MEMORY_MANAGEMENT
1229 /*
1230 ** This function is called to free superfluous dynamically allocated memory
1231 ** held by the pager system. Memory in use by any SQLite pager allocated
1232 ** by the current thread may be sqlite3_free()ed.
1233 **
1234 ** nReq is the number of bytes of memory required. Once this much has
1235 ** been released, the function returns. The return value is the total number
1236 ** of bytes of memory released.
1237 */
1238 int sqlite3PcacheReleaseMemory(int nReq){
1239   int nFree = 0;
1240   assert( sqlite3_mutex_notheld(pcache1.grp.mutex) );
1241   assert( sqlite3_mutex_notheld(pcache1.mutex) );
1242   if( sqlite3GlobalConfig.pPage==0 ){
1243     PgHdr1 *p;
1244     pcache1EnterMutex(&pcache1.grp);
1245     while( (nReq<0 || nFree<nReq)
1246        &&  (p=pcache1.grp.lru.pLruPrev)!=0
1247        &&  p->isAnchor==0
1248     ){
1249       nFree += pcache1MemSize(p->page.pBuf);
1250       assert( PAGE_IS_UNPINNED(p) );
1251       pcache1PinPage(p);
1252       pcache1RemoveFromHash(p, 1);
1253     }
1254     pcache1LeaveMutex(&pcache1.grp);
1255   }
1256   return nFree;
1257 }
1258 #endif /* SQLITE_ENABLE_MEMORY_MANAGEMENT */
1259
1260 #ifdef SQLITE_TEST
1261 /*
1262 ** This function is used by test procedures to inspect the internal state
1263 ** of the global cache.
1264 */
1265 void sqlite3PcacheStats(
1266   int *pnCurrent,      /* OUT: Total number of pages cached */
1267   int *pnMax,          /* OUT: Global maximum cache size */
1268   int *pnMin,          /* OUT: Sum of PCache1.nMin for purgeable caches */
1269   int *pnRecyclable    /* OUT: Total number of pages available for recycling */
1270 ){
1271   PgHdr1 *p;
1272   int nRecyclable = 0;
1273   for(p=pcache1.grp.lru.pLruNext; p && !p->isAnchor; p=p->pLruNext){
1274     assert( PAGE_IS_UNPINNED(p) );
1275     nRecyclable++;
1276   }
1277   *pnCurrent = pcache1.grp.nPurgeable;
1278   *pnMax = (int)pcache1.grp.nMaxPage;
1279   *pnMin = (int)pcache1.grp.nMinPage;
1280   *pnRecyclable = nRecyclable;
1281 }
1282 #endif